The patrol changes what the system gets to see
A predictive-policing system may treat past arrests as a signal of where crime is likely to happen. But an arrest record is not a direct measure of all crime. It records an encounter that police observed and acted on, which depends partly on where officers were sent and what they looked for.
That creates a feedback loop. If a model flags a neighborhood, officers may spend more time there. More patrol time means more chances to notice certain offenses, make stops, and record arrests. When those records are later used to update the model, the same area can look even more crime-prone—not necessarily because crime increased, but because it received more attention.
More data can reinforce a skewed picture
The process can make a prediction appear self-confirming: the system sends officers to a place, their activity generates the data, and that data supports sending them back. Meanwhile, offenses in less-patrolled areas may be less likely to enter the records the model learns from.
This does not mean every arrest is mistaken or that every flagged location is safe. It means arrest data mixes two things: where offenses occur and where police activity makes offenses more likely to be detected. If a system cannot distinguish those influences, adding fresh data may repeat the pattern rather than correct it.
The key question is how the data was made
Evaluating such a system therefore requires more than checking whether its forecasts match its own arrest records. It also requires asking how patrol decisions shaped those records, and whether the data captures underlying harm or simply past enforcement. Otherwise, a prediction can become part of the reason it seems accurate.
